breaking API:keras output_loss, shape of circuit wavefunction;some refactor on application codebase and reactivate some example demos into CI

Author: refraction-ray

.github/workflows/ci.yml

@@ -36,12 +36,14 @@ jobs:
         run: |
           cd docs
           make html
-      # - name: run example demos
-      #   run: |
-      #     cd examples
-      #     python GHZ_DQAS.py
-      #     python QEM_DQAS.py
-      #     python QAOA_DQAS.py
+      - name: run example demos
+        run: |
+          cd examples
+          python mcnoise_check.py
+          python vqnhe_h6.py
+          python qaoa_dqas.py
+      #   python GHZ_DQAS.py
+      #   python QEM_DQAS.py
       - name: setup build
         run: |
           python3 setup.py build

examples/H6_hamiltonian.npy

@@ -1,25 +1,13 @@
-import cirq
-import numpy as np
-import scipy
-import sympy as sy
-import networkx as nx
-import tensorflow_quantum as tfq
-import tensorflow as tf
 from functools import partial
-from functools import lru_cache
-import tensornetwork as tn
-from itertools import product
-import json
 import sys
 
 sys.path.insert(0, "../")
+
+import numpy as np
+import tensorflow as tf
+
 import tensorcircuit as tc
-from tensorcircuit.applications.layers import *
-from tensorcircuit.applications.van import *
-from tensorcircuit.applications.graphdata import *
-from tensorcircuit.applications.dqas import *
-from tensorcircuit.applications.vags import *
-from tensorcircuit.applications.vqes import *
+from tensorcircuit.applications.vqes import VQNHE, JointSchedule
 
 tc.set_backend("tensorflow")
 tc.set_dtype("complex128")
@@ -34,8 +22,8 @@ def initial_param(t, last=None, lastlast=None):
             qw.numpy() + np.random.uniform(low=-0.1, high=0.1, size=qw.numpy().shape)
         )
         cw = last[-2]
-        for i, t in enumerate(cw):
-            cw[i] = t + np.random.uniform(low=-0.1, high=0.1, size=t.shape)
+        for i, b in enumerate(cw):
+            cw[i] = b + np.random.uniform(low=-0.1, high=0.1, size=t.shape)
         return {"c": cw, "q": qw}
 
     return {}
@@ -106,7 +94,7 @@ def learn_c():
 if __name__ == "__main__":
     history = []
     lihh = np.load("data_file")
-    for i, h in enumerate(lihh[3:6]):
+    for h in lihh[3:6]:
         history.append(adiabatic_range(h.tolist(), history))
     print(history)
     # vqeinstance = VQNHE(

examples/adiabatic_vqnhe.py

@@ -11,7 +11,7 @@
 
 n = 5
 nlayer = 3
-mctries = 100000
+mctries = 100  # 100000
 
 print(jax.devices())
 
@@ -59,7 +59,12 @@ def f(key):
 for i in tqdm(range(mctries)):
     key, subkey = jax.random.split(key)
     psi = f(subkey)  # [1, 2**n]
-    rho += 1 / mctries * tc.backend.transpose(psi) @ tc.backend.conj(psi)
+    rho += (
+        1
+        / mctries
+        * tc.backend.reshape(psi, [-1, 1])
+        @ tc.backend.conj(tc.backend.reshape(psi, [1, -1]))
+    )
 
 print(rho)
 print("difference\n", tc.backend.abs(rho - rho0))

examples/mcnoise_check.py

@@ -2,8 +2,6 @@
 
 sys.path.insert(0, "../")
 
-import numpy as np
-import tensorflow as tf
 import tensorcircuit as tc
 from tensorcircuit.applications.dqas import (
     parallel_qaoa_train,
@@ -16,7 +14,10 @@
 tc.set_backend("tensorflow")
 
 set_op_pool([Hlayer, rxlayer, rylayer, rzlayer, xxlayer, yylayer, zzlayer])
+
+
 if __name__ == "__main__":
+    # old fashion example, prefer vvag for new generation software for this task
     parallel_qaoa_train(
         [0, 6, 1, 6, 1],
         single_generator(get_graph("8B")),

examples/qaoa_parallel_opt.py

@@ -1,30 +1,14 @@
-import cirq
 import numpy as np
-import scipy
-import sympy as sy
-import networkx as nx
-import tensorflow_quantum as tfq
-import tensorflow as tf
-from functools import partial
-from functools import lru_cache
-import tensornetwork as tn
-from itertools import product
-import json
 import sys
 
 sys.path.insert(0, "../")
 import tensorcircuit as tc
-from tensorcircuit.applications.layers import *
-from tensorcircuit.applications.van import *
-from tensorcircuit.applications.graphdata import *
-from tensorcircuit.applications.dqas import *
-from tensorcircuit.applications.vags import *
-from tensorcircuit.applications.vqes import *
+from tensorcircuit.applications.vqes import VQNHE, JointSchedule
 
 tc.set_backend("tensorflow")
 tc.set_dtype("complex128")
 
-h6h = np.load("/PATH")  # reported in 0.99 A
+h6h = np.load("./H6_hamiltonian.npy")  # reported in 0.99 A
 
 vqeinstance = VQNHE(
     10,
@@ -41,14 +25,13 @@ def learn_q():
 
 
 def learn_c():
-    #     return 0.
     lr = np.random.choice([0.004, 0.006, 0.008, 0.1])
     return JointSchedule(200, 0.0006, 10000, lr, 5000)
 
 
 rs = vqeinstance.multi_training(
-    tries=10,
-    maxiter=15000,  # 10000
+    tries=2,  # 10
+    maxiter=500,  # 10000
     threshold=0.5e-8,
     learn_q=learn_q,
     learn_c=learn_c,

examples/vqnhe_h6.py

@@ -6,7 +6,6 @@
 from functools import partial
 from typing import Any, Dict, Iterator, Sequence, Tuple, Optional
 
-import cirq
 import networkx as nx
 import numpy as np
 
@@ -468,6 +467,9 @@ def two2one(x: int, y: int) -> int:
 
 
 def dress_graph_with_cirq_qubit(g: Graph) -> Graph:
+    # deprecated cirq related utilities
+    import cirq
+
     for i in range(len(g.nodes)):
         g.nodes[i]["qubit"] = cirq.GridQubit(i, 0)
     return g

tensorcircuit/applications/graphdata.py

@@ -166,6 +166,7 @@ def f(circuit: Circuit, symbol: Union[Tensor, float], g: Graph = None) -> Circui
 
 
 def generate_double_gate_layer_bitflip(gates: str) -> None:
+    # deprecated, as API are consistent now for DMCircuit and Circuit
     def f(
         circuit: DMCircuit, symbol: Union[Tensor, float], g: Graph, *params: float
     ) -> DMCircuit:
@@ -259,7 +260,7 @@ def f(circuit: Circuit, symbol: Tensor, g: Graph = None) -> Circuit:
 def anyswaplayer(circuit: Circuit, symbol: Tensor, g: Graph) -> Circuit:
     for i, e in enumerate(g.edges):
         qubit1, qubit2 = e
-        circuit.exp(  # type: ignore
+        circuit.exp1(  # type: ignore
             qubit1,
             qubit2,
             unitary=array_to_tensor(_swap_matrix),
@@ -274,7 +275,7 @@ def anyswaplayer_bitflip_mc(
 ) -> Circuit:
     for i, e in enumerate(g.edges):
         qubit1, qubit2 = e
-        circuit.exp(  # type: ignore
+        circuit.exp1(  # type: ignore
             qubit1,
             qubit2,
             unitary=array_to_tensor(_swap_matrix),
@@ -322,7 +323,8 @@ def bitfliplayer_mc(ci: Circuit, g: Graph, px: float, py: float, pz: float) -> N
 
 
 ## below is similar layer but in cirq API instead of tensrocircuit native API
-## special notes to the API, the arguments order are different due to historical reason
+## special notes to the API, the arguments order are different due to historical reason compared to tc layers API
+## and we have no attention to further maintain the cirq codebase below, availability is not guaranteend
 
 basis_rotation = {
     "x": (cirq.H, cirq.H),

tensorcircuit/applications/layers.py

@@ -377,7 +377,7 @@ def exp_forward(
             layer, graph = cset[j]
             layer(ci, theta[i], graph)
 
-    state = ci.wavefunction()[0]
+    state = ci.wavefunction()
     losses = ave_func(state, g, *fs)
     return losses
 

tensorcircuit/applications/vags.py

@@ -103,9 +103,8 @@ def _build_masks(
                     assert (
                         hidden_space % (input_space - 1) == 0
                     ), "hidden space must be multiple of input space -1 when you set evenly as True"
-                    m = np.arange(
-                        1, input_space
-                    )  # TODO: whether 0 is ok need further scrunity, this is a bit away from original MADE idea
+                    m = np.arange(1, input_space)
+                    # TODO: whether 0 is ok need further scrunity, this is a bit away from original MADE idea
                     self._m.append(
                         np.hstack([m for _ in range(hidden_space // (input_space - 1))])
                     )
@@ -177,9 +176,8 @@ def call(self, inputs):
         if self.probamp is not None:
             inputs = self.probamp + inputs
         if self.nonmerge:
-            inputs = (
-                nonmerge_block + inputs
-            )  # TODO: to be investigated whether the nonmerge block can affect on unity of probability
+            inputs = nonmerge_block + inputs
+            # TODO: to be investigated whether the nonmerge block can affect on unity of probability
         outputs = tf.keras.layers.Softmax()(inputs)
         return outputs